WO2011098752A2 - Analyse d'image - Google Patents

Analyse d'image Download PDF

Info

Publication number
WO2011098752A2
WO2011098752A2 PCT/GB2011/000157 GB2011000157W WO2011098752A2 WO 2011098752 A2 WO2011098752 A2 WO 2011098752A2 GB 2011000157 W GB2011000157 W GB 2011000157W WO 2011098752 A2 WO2011098752 A2 WO 2011098752A2
Authority
WO
WIPO (PCT)
Prior art keywords
models
image
interest
model
region
Prior art date
Application number
PCT/GB2011/000157
Other languages
English (en)
Other versions
WO2011098752A3 (fr
Inventor
Graham Richard Vincent
Michael Antony Bowes
Gwenael Alain Guillard
Ian Michael Scott
Original Assignee
Imorphics Limited
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Priority claimed from US12/703,438 external-priority patent/US9530077B2/en
Priority claimed from GBGB1002270.5A external-priority patent/GB201002270D0/en
Application filed by Imorphics Limited filed Critical Imorphics Limited
Priority to JP2012552455A priority Critical patent/JP5955782B2/ja
Priority to EP11704473.5A priority patent/EP2534613B1/fr
Publication of WO2011098752A2 publication Critical patent/WO2011098752A2/fr
Publication of WO2011098752A3 publication Critical patent/WO2011098752A3/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
    • G06V10/00—Arrangements for image or video recognition or understanding
    • G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
    • G06V10/74—Image or video pattern matching; Proximity measures in feature spaces
    • G06V10/75—Organisation of the matching processes, e.g. simultaneous or sequential comparisons of image or video features; Coarse-fine approaches, e.g. multi-scale approaches; using context analysis; Selection of dictionaries
    • G06V10/755—Deformable models or variational models, e.g. snakes or active contours
    • G06V10/7557—Deformable models or variational models, e.g. snakes or active contours based on appearance, e.g. active appearance models [AAM]

Definitions

  • the present invention relates to methods and apparatus for identifying features in an image. More particularly, but not exclusively, the present invention relates to methods and apparatus for fitting models to an image to identify features in the image.
  • Automated image analysis may be used to reduce manual workloads in determining properties of images.
  • Image analysis is used in a variety of different fields to identify features of images.
  • One area in which image analysis techniques are used is to process medical images where the identification of anatomical features in an image can provide useful information to medical practitioners. Often many images are required to be processed, for example in a screening program, and each image should be processed as accurately as possible. Images of a particular anatomical feature obtained from different patients vary considerably. This may particularly be the case when an anatomical feature is in a diseased state. It can therefore be difficult to accurately identify anatomical features in images given such variance.
  • An existing approach to identifying anatomical features in an image is based upon statistical modelling of an anatomical region of interest.
  • Statistical models are generated from a training set of images in which features within the anatomical region of interest have been manually identified and a statistical model indicates a mean model and an acceptable degree of variation from the mean model, as indicated by the training set. Where a small number of training images are available, a statistical model can often fail to identify particular anatomical features in previously unseen images due to the failure of the statistical model to capture a true indication of acceptable variation.
  • a computer-implemented method for identifying features in an image comprises fitting a plurality of second models to the image, the plurality of second models together modelling a region of interest, wherein each part of the region of interest is modelled by at least two of the plurality of second models; and identifying the features in the image based upon the fit of the plurality of second models.
  • Fitting a plurality of second models to an image allows localised regions to be individually modelled.
  • the local nature of the second models can allow a more accurate fit to the image.
  • By including parts of the region of interest to be modelled in more than one of the second models if the fit of any one of the second models to a particular region is unacceptable, another of the second models that models the particular region can be used.
  • the method may further comprise initially fitting a first model of the object to the image and fitting the plurality of second models may be based upon the fitting of the first model.
  • Fitting the plurality of second models to the image may be initialised based upon the fitting of the first model. Initialising the fitting of models based upon a fit of a first model or models in this way allows information determined from fitting a first model to guide a fitting of subsequent models. Alternatively or additionally fitting the plurality of second models may be limited by the fitting of the first model.
  • Identifying features in the image based upon the fit of the plurality of second models may comprise combining data based upon the fit of the plurality of second models.
  • Combining data based upon the fit of the plurality of second models may comprise, for each part of the region of interest included in at least two of the plurality of models, applying an associated weight to data derived from each of the plurality of second models in which the part is included.
  • the associated weight may be based upon a distance of the part of the region from a point defined with respect to the particular second model.
  • the point defined with respect to the particular second model may be defined with respect to a centre of the particular second model, for example the centre of a region modelled by a particular second model.
  • a particular feature may be modelled by two of the second models and the combining may identify the particular feature by combining data associated with the particular feature generated by fitting each of the two of the second models.
  • the combining may be based upon stored data indicating a plurality of relationships, each relationship being between at least part of a second model and at least part of at least one other second model.
  • each point in a particular second model may have a relationship with any point in any other second model that represents a corresponding point in a region of interest.
  • the stored data indicating a plurality of relationships may indicate a relationship between at least part of a second model and part of the first model.
  • the combining may comprise, for each part of the region of interest included in at least two of the plurality of second models, determining a weighted average fit for the part of the region of interest based upon the fit of each of the at least two of the plurality of second models to the part of the region of interest.
  • the predetermined criterion may be a relationship between the initial estimate of location of the features and the fit of the one of the second models.
  • the predetermined criterion may comprise a geometric criterion.
  • the predetermined criterion may be determined for the one of the second models during a training phase, based upon a set of training images.
  • the predetermined criterion may be based upon a plurality of data items, each data item representing a fit of the one of the second models to a respective one of the training images.
  • Each of the plurality of second models may have an associated predetermined criterion.
  • the method may further comprise fitting a plurality of third models to the image, the plurality of third models together modelling at least part of the region of interest, the fitting the plurality of third models being based upon the estimate of location of the features, and identifying the features in the image based upon the fit of the plurality of third models.
  • the estimate of location of the features may be used to initialise the third models in the image.
  • the estimate of location of the features may be updated based upon the fit of the plurality of third models.
  • Features in the image may be identified based upon the updated estimate of location of features.
  • Further pluralities of models may be fitted to the image based upon the estimate of location of the features, with the estimate of location of the features updated after fitting each plurality of models.
  • a given part of the region of interest may be modelled by a first number of the second models and a second different number of the third models.
  • the second number may be greater than the first number.
  • Models in the third plurality of models may generally model smaller regions than models in the second plurality of models. In this way increasingly localised regions of the image may be modelled and a more accurate fit to the image may be achieved.
  • the method may further comprise generating the plurality of second models.
  • a training set of images may be received, each image in the training set of images representing the region of interest, and generating the plurality of second models of the object may based upon the training set of images.
  • Each model in the plurality of second models may be generated based upon a subset of images in the training set of images, the subset being selected by randomly sampling with replacement from the training set of images.
  • Generating the plurality of second models may comprise generating a plurality of models, each of the plurality of models modelling at least part of the region of interest; processing the plurality of models with respect to a further training set of images; and selecting a subset of the plurality of models based upon the processing such that each part of the region of interest is represented by a plurality of models in the subset of models.
  • Generating the plurality of second models may further comprise generating a first plurality of second models such that each part of the region of interest is modelled by a plurality of the first plurality of second models; generating a second plurality of second models such that each part of the region of interest is modelled by a plurality of the second plurality of second models; processing the first plurality of second models and the second plurality of second models with respect to the further training set of images; and selecting one of the first plurality and the second plurality of second models of the region of interest based upon the processing.
  • Processing the first plurality of second models and the second plurality of second models with respect to a training set of images may comprise fitting the first plurality of second models to each image in the further training set of images; and fitting the second plurality of second models to each image in the further training set of images; and selecting one of the plurality of second models of the region of interest based upon the processing may be based upon the fittings.
  • Generating the plurality of second models of the object based upon the training set of images may comprise, for each of the plurality of points, processing a corresponding part of the region of interest in each image in the training set of images.
  • a computer- implemented method for identifying features in an image comprises fitting a plurality of second models to the image, the plurality of second models together modelling a region of interest; fitting a plurality of third models to the image, the plurality of third models together modelling at least part of the region of interest, the fitting of the plurality of third models being based upon the fit of the plurality of second models; and identifying the features in the image based upon the fit of the plurality of third models.
  • a fit to the image in a given region may be modified based upon fits of different sets of models.
  • Such a method provides a way of refining identification of an object in an image. Further pluralities of models may be fitted to the image based upon previous pluralities of models to further refine the identification.
  • a given part of the region of interest may be modelled by a first number of second models and a second different number of third models.
  • the second number may be greater than said first number.
  • Some or all of the third number of models may model a smaller region than some or all of the second number of models and so each of the third plurality of models may model a more localised part of the region of interest. Improved modelling of the region of interest may thereby be achieved.
  • the method may further comprise initially fitting a first model of the region of interest to the image and fitting the plurality of second models may be based upon the fitting of the first model. Fitting the plurality of second models to the image may be initialised based upon the fitting of the first model and fitting the plurality of third models to the image may be initialised based upon the fitting of the first model. As indicated above, initialising models based upon a fit of previous models in this way allows information determined from fitting the previous models to guide the subsequent fitting.
  • Each region of the object may be represented by a plurality of the plurality of second models. Advantages described for the first aspect of the invention may therefore also be achieved in the second aspect of the invention.
  • a computer-implemented method for identifying features in an image comprises determining an estimate of location of the features in the image; fitting a second model to part of the image; and updating the estimate based upon the second model if the fit of the second model satisfies a predetermined criterion.
  • the predetermined criterion may be based upon a difference between the estimate and the fit of the second model.
  • the predetermined criterion may be a relationship between the estimate of location of the features and the fit of the second model.
  • the predetermined criterion may comprise a geometric criterion, for example a maximum distance between the points modelled by the second model after fitting and corresponding points in the estimate of location of the features or a maximum amount of rotation of part of the image modelled by the second model and a corresponding part of the estimate of location, or a combination of both.
  • the predetermined criterion may be determined for the second model during a training phase, based upon a set of training images.
  • the predetermined criterion may be based on a plurality of data items, each data item representing a fit of the second model to a respective one of said training images.
  • the second model may be fitted to a plurality of training images and the fit of the second model to each training image may be classified as a "good” or "bad” fit.
  • the classified fits may be used to train a classifier for determining whether a particular fit of a second model is an acceptable fit.
  • the method may further comprise fitting a first model to the image and the estimate of location of the features in the image may be based upon the fit of the first model. Fitting the second model to part of the image may be based upon the estimate of location of the features. Additionally, fitting the second model to part of the image may be initialised based upon the estimate of location of the features. Fitting a second model to part of the image may comprise fitting a plurality of second models to the image, and each of the plurality of second models may have an associated predetermined criterion. That is, a criterion for a particular second model may allow the particular second model to vary a relatively large amount from a starting position compared to a different second model.
  • a computer-implemented method of generating a plurality of second models that together model a region of interest comprises generating the plurality of second models such that each part of the region of interest is modelled by a plurality of the plurality of second models.
  • regions of a region of interest to be modelled in more than one of the second models if the fit of any one of the second models to a particular region is unacceptable, another of the second models that models the particular region can be used when the second models are used to model the region of interest.
  • the method may further comprise receiving a training set of images, each image in the training set of images representing the region of interest and generating the plurality of second models of the object may be based upon the training set of images.
  • Each model in the plurality of second models may be generated based upon a subset of images in the training set of images, the subset being selected by randomly sampling with replacement from the training set of images.
  • Generating the plurality of second models may comprise: generating a plurality of models, each of said plurality of models modelling at least part of the region of interest; processing the plurality of models with respect to a further training set of images and selecting a subset of the plurality of models based upon said processing such that each part of the region of interest is represented by a plurality of models in the subset of models.
  • Generating the plurality of second models may comprise: generating a first plurality of second models such that each part of the region of interest is modelled by a plurality of the first plurality of second models; generating a second plurality of second models such that each part of the region of interest is modelled by a plurality of the second plurality of second models; processing the first plurality of second models and the second plurality of second models with respect to the further training set of images; and selecting one of the first plurality of second models and second plurality of second models of the region of interest based upon the processing.
  • Processing the first plurality of second models and the second plurality of second models with respect to a training set of images may comprise: fitting the first plurality of second models to the object; and fitting the second plurality of second models to the object; and selecting one of said plurality of second models of the object based upon the processing may be based upon the fittings.
  • Generating the or each plurality of second models may comprise receiving data indicating a plurality of points of interest within the region of interest and generating the or each plurality of second models based upon the received plurality of points of interest.
  • the points of interest may be selected in any convenient way.
  • each point of interest may be indicative of a feature of interest such as, for example, features of the region of interest and/or regions known to be of greatest curvature in the region of interest. In this way, particular features which it is particularly desirable to accurately model can be selected so as to fall at the centre of a model, and therefore be more accurately modelled.
  • Generating the or each plurality of second models based upon the received data may comprise, for each of the received points of interest, selecting a part of the region of interest of a predetermined size to provide a basis for one of the plurality of models. Each part of the region of a predetermined size may be centred on one of the received points of interest.
  • Generating the plurality of second models of the object based upon the training set of images may comprise, for each of the plurality of points, processing a corresponding part of the region of interest in each image in the training set of images.
  • the models may be any suitable description of shape and/or image characteristics which can be fitted to an image.
  • the models may be rigid models that may be fitted to an image by changing the position and/or rotation of the models to determine a best fit to the image.
  • the models may be deformable models.
  • Deformable models are models in which shape of a region may be changed within predetermined constraints in order to effectively fit the models to the image. Examples of deformable models are statistical models such as shape or appearance models, deformable templates and deformable atlases. Where the models are appearance models they may represent a volume or a surface.
  • the or each image may be an image selected from the group consisting of: a computed topography image, a magnetic resonance image, a radiograph image, an x-ray image and an ultrasound image.
  • the fitting may comprise applying a fitting technique selected from the group consisting of: rigid registration, non-rigid registration, active shape modelling and active appearance modelling.
  • the plurality of second models may represent an arbitrary decomposition of the object.
  • the region of interest may be a part of a body.
  • features is used herein to indicate any part of an image which is to be identified.
  • a feature may be an image region, an image boundary, an image surface or a point in an image
  • the features may be anatomical features.
  • anatomical image regions include a region indicative of bone such as a femur, a region indicative of an organ or part of an organ (for example a region indicative of the frontal lobe), regions indicative of abnormalities (for example regions indicative of cysts or regions indicative of pathological bone growth).
  • anatomical boundaries and surfaces include bone/cartilage boundaries and/or surfaces and ventricle/cerebrospinal fluid boundaries and/or surfaces.
  • An example of an anatomical point in an image is an iliac crest.
  • aspects of the invention can be implemented in any convenient form. For example computer programs may be provided to carry out the methods described herein. Such computer programs may be carried on appropriate computer readable media which term includes appropriate tangible storage devices (e.g. discs). Aspects of the invention can also be implemented by way of appropriately programmed computers.
  • Figure 1 is a schematic illustration of a system for analysis of images according to an embodiment of the present invention
  • Figure 1A is a schematic illustration showing a computer of the system of Figure 1 in further detail
  • Figure 2 is an image suitable for processing using the system of Figure 1 ;
  • Figure 3 is a flowchart showing processing carried out to analyse an image in the system of Figure 1 ;
  • Figure 4 shows the image of Figure 2 divided into parts, each of which is modelled by a respective model;
  • Figure 5 is a flowchart showing part of the processing of Figure 3 in further detail
  • Figure 6 is a flowchart showing part of the processing of Figure 5 in further detail
  • Figure 7 is a graph illustrating a relationship between a weight applied to a point and a distance of the point from the centre of a model.
  • a computer 1 is arranged to receive an image 2 of an object or objects and to receive a model 3.
  • the model 3 is fitted to the image 2 to generate a parameterisation of the model 4.
  • the model 3 is a statistical model such as an active appearance model or active shape model, which has been created based upon a training set of images of an object of the same class as the object that is the subject of image 2.
  • the model 3 is a statistical model of the variation within the training images and is generated by processing the training set of images to generate a mean model and a range of variation from the mean model of the training set.
  • the parameterisation 4 generated by computer 1 can be used to identify features of the image.
  • Figure 1 A shows the computer 1 in further detail. It can be seen that the computer comprises a CPU l a which is configured to read and execute instructions stored in a volatile memory lb which takes the form of a random access memory.
  • the volatile memory lb stores instructions for execution by the CPU la and data used by those instructions.
  • the image 2 and model 3 may be stored in the volatile memory lb.
  • the Computer 1 further comprises non-volatile storage in the form of a hard disc drive l c.
  • the image 2 and model 3 may be stored on the hard disc drive l c.
  • the computer 1 further comprises an I/O interface Id to which is connected peripheral devices used in connection with the computer 1. More particularly, a display le is configured so as to display output from the computer 1. The display le may, for example, display a representation of the image 2. Additionally, the display l e may provide data representing the parameterisation 4 to a user and may display an input image 2 overlaid with the parameterisation 4. Input devices are also connected to the I/O interface Id. Such input devices include a keyboard l e and a mouse I f which allow user interaction with the computer 1.
  • a network interface 1 g allows the computer 1 to be connected to an appropriate computer network so as to receive and transmit data from and to other computing devices.
  • the CPU la, volatile memory lb, hard disc drive l c, I/O interface Id, and network interface lg, are connected together by a bus lh.
  • FIG 2 an image suitable for processing by the system of Figure 1 is shown.
  • Figure 2 is a slice through a Magnetic Resonance (MR) volume image of the knee.
  • MR Magnetic Resonance
  • the processing described below makes reference to a model Mi, built based upon a training set of images.
  • the model Mi is a statistical model built based upon a training set of images and is a model of features of interest in an image I, as set out above.
  • generation of the model Mi and fitting of model Mi to an image I will first be described. It will be appreciated however that the generation and fitting of the model Mi can be carried out in any convenient way.
  • the model Mi can be generated using a parameter! sation technique described in United States Patent Number 7584080, the contents of which are herein incorporated by reference.
  • the model Mi is generated from a training set of images by first manually marking corresponding predetermined points of interest in each of the images in the training set of images.
  • the set of points of interest are selected based upon features of the object to be modelled. For example, if the object to be modelled is a knee joint, points of interest may be selected based upon anatomical features of the knee joint.
  • the model Mi may be, for example, a shape model or an appearance model, which are both described below. It will be appreciated, however, that other models may be used such as for example, deformable templates or deformable atlases which take the form of a typical or average example which is allowed to deform within a set of constraints.
  • model Mi is a shape model
  • the set of points of interest associated with each training image, selected in the manner described above is represented by a shape vector comprising a concatenation of the coordinates of the selected points.
  • the shape vector x for a particular training image takes the form:
  • X (Wn 2 --, * makeup,. ,, ) (!) where n is the number of points of interest in each image.
  • a mean sha e vector x can be calculated according to equation (2):
  • N is the number of training images
  • Xi is a vector representing a training image i
  • a plurality of shape vectors, one for each training image can be used to generate a matrix X, in which each row represents a shape vector.
  • Each of the shapes represented by the points selected in a respective training image is translated, rotated and scaled so that their centroids coincide and the sum of squared differences between the points in each image and those on the mean shape is minimised.
  • the eigenvectors of the matrix S can be determined and a matrix P of the t most significant eigenvectors can be created.
  • x is a mean shape vector generated from shape vectors of images in the training set (according to equation (2));
  • P s is the matrix of eigenvectors described above.
  • b s is a t-element vector of weights to be applied to eigenvectors of the matrix P s to generate the vector x for the shape.
  • the eigenvectors Pj are sorted according to their eigenvalues ⁇ ⁇ , which describe the observed variance over the training shapes associated with each eigenvector. That is to say the ⁇ 3 ⁇ 4 represent the proportions of variance from the mean shape that are described by each of the eigenvectors.
  • t ⁇ N
  • the model can be made to represent some chosen proportion of the total variance observed in the training shapes.
  • Each eigenvector included in the matrix P s represents a mode of variation of the modelled shape. Varying values of elements of the vector b s allows generation of shapes within the range observed within images of the training images.
  • An appearance model is an extension of the shape model described above.
  • grey-level information gj m is sampled from the image over the region covered by the mean shape x , after normalising the image to minimise the effect of global lighting variation.
  • g is the mean normalised grey-level vector
  • P g is a matrix of eigenvectors
  • b g is a set of grey-level parameters.
  • Varying values of elements of the vectors b s (in equation (5)) and b g (in equation (6)) allows generation of shapes and grey-level images within the range observed within images of the training images.
  • W s is a diagonal matrix of weights associated with each shape parameter in the vector b s .
  • Equation (8) The matrix W s adjusts values of b s to allow for the difference in units between the shape and grey level models.
  • a further model shown in equation (8) is generated from the set of concatenated vectors b, using the method described above to generate models (5) and (6).
  • b Qc (8) where:
  • Q is a matrix of eigenvectors
  • c is a vector of appearance parameters.
  • An example image can be synthesised for a given c by generating the shape-free grey- level image from the vector g and warping it using the control points described by x.
  • the aim of model fitting is to determine the set of parameters c which best fit the model to an unseen image.
  • One way to do this is by using active appearance models which are described below.
  • An active appearance model uses the difference between a reconstructed image generated by the model and the underlying target image I, to iteratively drive the model parameters towards better values.
  • a model Mi is fitted to the image I.
  • the model Mi is a statistical model built based upon a training set of images and is a model of features of interest in the image I such as the models described above.
  • a counter k is set to a value 2 and at step S4 each model m of a set of models M k is initialised in the image I based upon the fit of the model Mi.
  • the set of models M 2 is a decomposition of an anatomical region modelled by the model Mi. That is, each point modelled by each model m corresponds to a point modelled by the model Mi, and collectively the set of models M 2 model all of the features of interest modelled by the model Mi in the image I. For all values of k , models of the set of models M 2 overlap such that each point of the object in the image 2 that is modelled by model Mi is modelled by a plurality of the models m in the set of models M 2 . Decomposition of the anatomical region modelled by the model Mi to generate a set of models M is described in further detail below with reference to Figure 4.
  • Initialisation of each model m is performed by fitting points in each model to the image I based upon the fit of corresponding points of the model M ⁇ , determined at step S2. Corresponding points are determined using a point correspondence indicating a relationship between points.
  • the point correspondence relates each point modelled by the model Mi to points in models m (for all M k ) that correspond to the respective point. In this way, the relationship between points in each model in a set of models M k is stored, as well as the relationship between points in models m k in a set of models M k and models ⁇ 3 ⁇ 4+ ⁇ in a set of models M k+1 .
  • each model m is adjusted based upon properties of the image I and properties of the model m in the same way as the fitting of model Mi described above.
  • Each model m is adjusted independently of any other model m in the set of models M k and each model m generates an estimate of location of features in the image I that are modelled by m.
  • the fits of the models in the set M k to the image I after the adjustment of step S5 are combined to generate a combined estimate which provides an estimate of the location of each of the features of interest in the image.
  • the combined estimate may take the form of any suitable data.
  • the combined estimate may be a geometric indication of the location of each of the features of interest in the image I, or may be a description of the image I indicating the features of interest in the image I, or a combination of both.
  • Combining the fits of the models M k to generate a combined estimate is described in further detail below with reference to Figure 5.
  • a check is performed to determine whether there are more sets of models M k to be fitted to the image. If it is determined that there are more sets of models then at step S8 the counter k is incremented and processing passes to step S4 where the further set of models M k+ i is initialised based upon the combined estimate, determined at step S6. Otherwise at step S9 an estimate of location of features of interest is output based upon the combined estimate generated from the set of models M k .
  • Each set M k+ i generally contains a greater number of models than the previous set M k and, in general, models in the set M k+ i will model a smaller number of points than models in the set M 3 ⁇ 4 . As such, as k increases models in the set M k will generally model a more localised region of the object and smaller regions of the object will generally be modelled with greater accuracy.
  • Initialisation of model fitting based upon the combined estimate minimises the risk of a model incorrectly identifying part of the image I as a feature of interest when such a part of the image I is, in fact, not a feature of interest. That is, where a model models a relatively small part of an image, there is a greater chance such a model will fit an incorrect part of an image well. As such, smaller models need better initialisation to avoid false fitting.
  • Use of a hierarchical scheme allows a plurality of models of differing sizes to be applied to an image, such that fitting of relatively large models initializes the fitting of relatively smaller models.
  • Models Mj +1 at each level of the hierarchy update an estimate generated by a level Mi of the hierarchy.
  • FIG. 4 illustrates a decomposition of an anatomical region modelled by a model Mi based upon the image of Figure 2.
  • Figure 4 shows three regions 10, 1 1, 12. Each of regions 10, 1 1 , 12 corresponds to part of the image of Figure 2. Each of regions 10, 1 1 , 12 corresponds to a subset of the points of interest modelled by model Mi and indicates regions of the image of Figure 2 that correspond to one of the models in the set of models Mk. It can be seen that regions 10, 1 1, 12 overlap so that some points fall within two or three of regions 10, 1 1 , 12.
  • image feature 13 of region 10 corresponds to features 14 and 15 of regions 1 1 and 12 respectively, such that any points corresponding to features 13, 14 and 15 that are modelled by model Mi will be modelled by each of the models corresponding to image regions 10, 1 1 and 12. It has been found that modelling each image point with between 5 and 15 models in the set of models M k is effective, although each image point may be modelled in other numbers of models greater than or equal to 2, in a set Mk-
  • each set of models M is determined by selecting a plurality of points of interest and identifying a region based upon each of the selected points.
  • the set of points selected to form the basis for a model m of the set of models M 3 ⁇ 4 is processed with respect to a training set of images to generate the model m in the manner described above.
  • the training set of images used can be the same training set as is used to generate the model Mi, or could be a further training set of images which have been parameterised with corresponding points to the points used in the generation of the model Mi.
  • a subset of the set of training images may be used to generate each model with each subset selected by randomly sampling with replacement from the set of training images.
  • each model may be generated based upon a randomly selected subset of the training images, wherein the subset selected for generation of a particular model does not affect subsets selected for generation of other models.
  • the particular points of interest modelled by any model m in a set of models M k may be selected in any convenient way. For example a segmenter may be used to divide the set of points of interest modelled by an initial model into overlapping subsets. The subsets may be determined by selecting points at random and determining an area of a predetermined size surrounding each selected point in each training image to provide basis for a model. That is, any points falling within the determined area may then be used as a basis for a model of the set of models M k .
  • Points may be chosen in the same way until a desired degree of overlap has been achieved for each of the points to be modelled.
  • a set of points may be selected that are evenly distributed across an initial model, such that the distance between the selected points is approximately equal.
  • An area surrounding each of the selected points may then be selected and any points of the initial model falling within the area associated with a selected point may be used as a basis for a model of the set of models M k . In this way, the amount of overlap can be controlled such that an even distribution of overlap is achieved.
  • the initial model may be processed based upon the mean model of the initial model. That is, points from the mean model of the initial model may be used as a basis for generation of further models.
  • a still further alternative is to select a set of points in an initial model associated with features of interest associated with particular image features. For example, points in regions having a high degree of curvature may be selected and for each selected point a model m may be determined as above, by selecting an area surrounding the selected point and including each point in the area in a model m.
  • the size of an area modelled by a model m + i is selected such that each of the models in a set of models M k +i models a smaller region of interest than a model i k in the corresponding set of models M k .
  • the size of an area modelled by a model may be chosen based upon, for example, the size of features that it is desired to be accurately modelled at stage k. As such, in some embodiments at least some models in a set of models M k+ i may not model a smaller region than models in a corresponding set of models M k . For example where certain features have been shown not to benefit from further decomposition, for example during a training stage, a region selected to model the features may remain relatively large throughout the various sets of models while in other areas relatively small regions may be selected to form the basis of a model.
  • the determination of a set of models M k is carried out in an initial stage, and as such a relatively large amount of time can be spent in selecting an effective set of models.
  • the above selection methods may each be carried out repeatedly and the efficacy of each set of models M k that is generated may be tested against images that have been manually marked to determine how well a particular set of models M k is able to represent variation in the test image data.
  • a set of models M k+ i generally contains a greater number of models than a set of models M k and that each model in the set of models M k+ i generally model a smaller region of an anatomical region than models in the set of models M k
  • each set of models M k is selected so as to effectively represent variation in the test image data.
  • a set of models M k+ i may contain the same or a smaller number of models than a set of models M k and/or models in the set of models M k+ ! may model regions of a similar size or a greater size than patches in a set of models M k where such choices are able to more effectively represent variation in the test image data.
  • a model m in the set of models M k is selected and at step SI 6 a check is performed to determine whether the fit of the selected model m is close enough to the fit of corresponding points of the combined estimate generated at step S6 for the set of models M k -i. Corresponding points between models and the combined estimate are determined by the point correspondence. Determination of whether the fit of a selected model m is close enough to the fit of corresponding points of the combined estimate generated from the set of models M k -i may be determined in any convenient way. For example a geometric rule may be determined for each region of the object to be modelled.
  • the geometric rule may indicate a maximum distance a model representing the region can move relative to the starting position of the model (as indicated by the combined estimate generated from the set of models M k -i) or a maximum number of degrees the model can rotate relative to the starting rotation of the model, or a combination of both. If the model does not exceed the maximum allowed, then the fit of the model is determined to be close enough.
  • a classification based rejection may be used in which the system is presented with a (preferably large) number of examples of a good fit of a model and a (preferably large) number of examples of a bad fit of a model in a training stage. In this way, the system is able to "learn" characteristics of a good fit and a bad fit based upon residual error between fit achieved by a model and ideal fit.
  • step SI 7 the model m is added to a set of acceptable models A k - If it is determined that the fit of the selected model m is not close enough at step S16 or if the model m has been added to A k at step SI 7, processing passes to step SI 8. At step SI 8 it is determined if there are previously unselected models in M k - If there are previously unselected models in M k then processing returns to step SI 5 where a previously unselected model is selected. Otherwise processing passes to step SI 9 where the models in the set A k are combined and used to update the combined estimate. Combination of models is described below with reference to Figures 6 and 7.
  • the processing of Figure 5 is based upon an assumption that, at each stage, the combined estimate is a reasonably accurate estimate of features in the image I. If after the adjustment of the set of models M k of step S5 points of a model m in the set of models M k have been adjusted by greater than a predetermined amount relative to corresponding points of the combined estimate then the model m is rejected as it is assumed that the model has moved too far, and is no longer modelling the correct part of the image. As described above, the set of models M k overlap such that each point of the region of interest indicated in the combined estimate is modelled by a plurality of the models in M k . Each point in the combined estimate can therefore still, in general, be adjusted based upon the set of models M k when a model m is rejected.
  • a point p in the combined estimate is selected and at step S26 a set of models M p is determined.
  • the set of models M p is a subset of the set of acceptable models A k which each model the point p, determined based upon the point correspondence.
  • a weight w m is determined for the point p for each model m in M p .
  • Each weight w m provides a weight for the contribution of a model m to the updating of point p in the combined estimate carried out at step S6 of Figure 3.
  • Each of the models in the set of models M p may indicate a different location for point p after the model has been adjusted and a weight w m determines how much influence a particular model m has upon the total adjustment of a point p in the combined estimate. Determination of a weight w m for a model m is described below with reference to Figure 7.
  • step S29 the point p in the combined estimate is updated based upon the weighted contributions of the point p in each of the models in the set M p . That is, for a particular model point p the fit of p is updated based upon data relating to that point in each model in the set of models M p .
  • the weights w therefore reflect the amount of information used to update a point p in any particular model m since where a point p lies relatively close to the centre of a region modelled by m c , a relatively large amount of information surrounding the point p has been modelled by m c and therefore used in the determination of the fit of the point p in the model m c . It can be seen that where a point p lies at the edge of a region modelled by a model m e , the weight w is small, but is greater than zero.
  • a model Mi is first fitted to an image I and a set of models M 2 , is subsequently fitted to the image I based upon the fit of the model Mi.
  • the fitting of the model Mi is avoided, and the combined estimate is first established by a set of models Mi. That is, in some embodiments a plurality of models are first fitted to the image I and combined to generate a combined estimate, and subsequent sets of models Mi +) are fitted to the image I in the manner described above.
  • a point correspondence indicates a relationship, for each point of interest p, between points in models m which model the point of interest p and a point in a combined estimate which corresponds to the point of interest p.
  • the combined estimate may be a label image in which each point in the combined estimate has a label indicating a structure or feature associated with the point.
  • each point in the combined estimate may have an associated label indicating that corresponding points in the image I are currently estimated to represent the femur. In this way, fitting a plurality of models to the image I and creation of the combined estimate can be carried out based upon labels.
  • each model that models a particular feature contributes to updating the combined estimate such that, after combination of the models, for each point indicated as representing a particular feature in image I, corresponding points in the combined estimate are labelled as representing the particular feature.

Landscapes

  • Engineering & Computer Science (AREA)
  • Software Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Theoretical Computer Science (AREA)
  • Computing Systems (AREA)
  • Physics & Mathematics (AREA)
  • Databases & Information Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • General Physics & Mathematics (AREA)
  • Multimedia (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)
  • Measuring And Recording Apparatus For Diagnosis (AREA)
  • Apparatus For Radiation Diagnosis (AREA)
  • Magnetic Resonance Imaging Apparatus (AREA)
  • Ultra Sonic Daignosis Equipment (AREA)

Abstract

L'invention porte sur un procédé mis en œuvre par ordinateur pour l'identification de caractéristiques dans une image. Le procédé comprend l'adaptation d'une pluralité de seconds modèles à l'image, la pluralité de seconds modèles modélisant ensemble une région d'intérêt, chaque partie de la région d'intérêt étant modélisée par au moins deux de la pluralité de seconds modèles ; et l'identification des caractéristiques dans l'image sur la base de l'adaptation de la pluralité de seconds modèles.
PCT/GB2011/000157 2010-02-10 2011-02-08 Analyse d'image WO2011098752A2 (fr)

Priority Applications (2)

Application Number Priority Date Filing Date Title
JP2012552455A JP5955782B2 (ja) 2010-02-10 2011-02-08 画像解析
EP11704473.5A EP2534613B1 (fr) 2010-02-10 2011-02-08 Analyse d'image

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
US12/703,438 US9530077B2 (en) 2010-02-10 2010-02-10 Image analysis
US12/703,438 2010-02-10
GBGB1002270.5A GB201002270D0 (en) 2010-02-11 2010-02-11 Image analysis
GB1002270.5 2010-02-11

Publications (2)

Publication Number Publication Date
WO2011098752A2 true WO2011098752A2 (fr) 2011-08-18
WO2011098752A3 WO2011098752A3 (fr) 2011-10-20

Family

ID=44269271

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/GB2011/000157 WO2011098752A2 (fr) 2010-02-10 2011-02-08 Analyse d'image

Country Status (3)

Country Link
EP (1) EP2534613B1 (fr)
JP (2) JP5955782B2 (fr)
WO (1) WO2011098752A2 (fr)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN110176024A (zh) * 2019-05-21 2019-08-27 腾讯科技（深圳）有限公司 在视频中对目标进行检测的方法、装置、设备和存储介质
WO2019211615A1 (fr) 2018-05-02 2019-11-07 Mako Surgical Corp. Segmentation d'image
WO2024006578A2 (fr) 2022-07-01 2024-01-04 Stryker Corporation Procédés et systèmes de planification de zone et d'implant pour une procédure chirurgicale

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7584080B2 (en) 2001-06-12 2009-09-01 Imorphics Limited Constructing a statistical shape model from two-dimensional or three-dimensional data

Family Cites Families (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP3533722B2 (ja) * 1994-09-30 2004-05-31 オムロン株式会社 不良検査方法およびその装置
JP4316114B2 (ja) * 2000-06-30 2009-08-19 コニカミノルタホールディングス株式会社 モデルの変形方法およびモデリング装置
JP3819236B2 (ja) * 2000-12-13 2006-09-06 株式会社東芝 パターン認識方法及びパターン認識を行わせるプログラムを記憶したコンピュータ読み取り可能な記憶媒体
DE10144004A1 (de) * 2001-09-07 2003-03-27 Philips Corp Intellectual Pty Verfahren zur Messung geometrischer Größen einer in einem Bild enthaltenen Struktur
US20060008143A1 (en) * 2002-10-16 2006-01-12 Roel Truyen Hierachical image segmentation
JP2005004612A (ja) * 2003-06-13 2005-01-06 Denso Corp 画像認識装置
US7536044B2 (en) * 2003-11-19 2009-05-19 Siemens Medical Solutions Usa, Inc. System and method for detecting and matching anatomical structures using appearance and shape
DE10357205A1 (de) * 2003-12-08 2005-07-14 Siemens Ag Verfahren zur Erzeugung von Ergebnis-Bildern eines Untersuchungsobjekts
JP2005198887A (ja) * 2004-01-16 2005-07-28 Fuji Photo Film Co Ltd 解剖学的構造物の構造検出方法および装置並びにプログラム、構造物除去画像生成装置、異常陰影検出装置
JP4701861B2 (ja) * 2005-06-22 2011-06-15 コニカミノルタエムジー株式会社 領域抽出装置、領域抽出方法およびプログラム
JP5017909B2 (ja) * 2005-06-22 2012-09-05 コニカミノルタエムジー株式会社 領域抽出装置、領域抽出方法及びプログラム
GB0521640D0 (en) * 2005-10-24 2005-11-30 Ccbr As Automatic quantification of a pathology indicating measure from cartilage scan data
GB0607143D0 (en) * 2006-04-08 2006-05-17 Univ Manchester Method of locating features of an object
US7988639B2 (en) * 2006-05-17 2011-08-02 St. Jude Medical, Atrial Fibrillation Division, Inc. System and method for complex geometry modeling of anatomy using multiple surface models

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US7584080B2 (en) 2001-06-12 2009-09-01 Imorphics Limited Constructing a statistical shape model from two-dimensional or three-dimensional data

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
"Deformable Models in Medical Image Analysis: A Survey", MEDICAL IMAGE ANALYSIS, vol. 1, no. 2, 1996, pages 91 - 108
T.F. COOTES; C.J. EDWARS; C.J. TAYLOR: "Active Appearance Models", IEEE TRANSACTIONS ON PATTERN ANALYSIS AND MACHINE INTELLIGENCE, vol. 23, no. 6, 2001

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2019211615A1 (fr) 2018-05-02 2019-11-07 Mako Surgical Corp. Segmentation d'image
US11017536B2 (en) 2018-05-02 2021-05-25 Mako Surgical Corp. Image segmentation
US11715208B2 (en) 2018-05-02 2023-08-01 Mako Surgical Corp. Image segmentation
CN110176024A (zh) * 2019-05-21 2019-08-27 腾讯科技（深圳）有限公司 在视频中对目标进行检测的方法、装置、设备和存储介质
CN110176024B (zh) * 2019-05-21 2023-06-02 腾讯科技（深圳）有限公司 在视频中对目标进行检测的方法、装置、设备和存储介质
WO2024006578A2 (fr) 2022-07-01 2024-01-04 Stryker Corporation Procédés et systèmes de planification de zone et d'implant pour une procédure chirurgicale

Also Published As

Publication number Publication date
JP5955782B2 (ja) 2016-07-20
EP2534613B1 (fr) 2019-03-27
EP2534613A2 (fr) 2012-12-19
JP2013519409A (ja) 2013-05-30
JP6275198B2 (ja) 2018-02-07
WO2011098752A3 (fr) 2011-10-20
JP2016170810A (ja) 2016-09-23

Similar Documents

Publication Publication Date Title
US10198662B2 (en) Image analysis
US8126291B2 (en) System and method for dense image registration using Markov Random Fields and efficient linear programming
US9824302B2 (en) Method and system for model-based fusion of multi-modal volumetric images
AU2018200223B2 (en) Method and node for manufacturing a surgical kit for cartilage repair
US9299145B2 (en) Image segmentation techniques
EP2289025A1 (fr) Procédés et système de détection et de suivi de l'ostéoporose
US8050469B2 (en) Automated measurement of objects using deformable models
JP2023162303A (ja) 整形外科的関節のための外科的な修正処置の術前計画
US11992348B2 (en) Machine learning model to adjust C-arm cone-beam computed tomography device trajectories
EP3107031A1 (fr) Procédé, appareil et système de marquage de la colonne vertébrale
AU2012293488B2 (en) Image processing method
Schmid et al. Segmentation of X-ray images by 3D-2D registration based on multibody physics
JP6275198B2 (ja) 画像解析
Morel-Forster et al. Probabilistic fitting of active shape models
JP2023505374A (ja) 医療画像セグメント化及びアトラス画像選択
US8867809B2 (en) Image processing method
Schramm et al. Toward fully automatic object detection and segmentation
US20220414869A1 (en) Detecting and segmenting regions of interest in biomedical images using neural networks
Cootes et al. Active shape and appearance models
Liu et al. Sparse intervertebral fence composition for 3D cervical vertebra segmentation
Romera et al. Multi-region statistical shape model for cochlear implantation
Pitiot et al. Expert knowledge guided segmentation system for brain MRI
Zheng et al. Reconstruction of 3D vertebral models from a single 2D lateral fluoroscopic image
WO2023017438A1 (fr) Système et procédé de translation d'image médicale
Huber et al. Texture feature selection with relevance learning to classify interstitial lung disease patterns

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 11704473

Country of ref document: EP

Kind code of ref document: A1

WWE Wipo information: entry into national phase

Ref document number: 2011704473

Country of ref document: EP

WWE Wipo information: entry into national phase

Ref document number: 2012552455

Country of ref document: JP

NENP Non-entry into the national phase

Ref country code: DE